Staged Training Report ✓ Complete

Run ID: shoulder_session_multiheight_vgg_loss_2
Generated: 2026-02-22 00:27:37
Stages Completed: 1
Total Elapsed Time: 05:31:39

Configuration

Config Defaults Changed Since Last Commit

Parameter	Previous	Current
`batch_size`	1	8
`lr_sweep.phase_b_time_budget_min`	10.0	5.0
`plateau_sweep.plateau_ema_alpha`	0.9	0.95
`plateau_sweep.plateau_improvement_threshold`	0.0005	0.005

All Configuration Parameters (60 parameters)

Parameter	Value
`total_samples`	10000000
`batch_size`	8
`stage_samples_multiplier`	100000000000
`update_interval`	250
`window_size`	100
`num_best_models_to_keep`	1
`sampling_mode`	Loss-weighted
`loss_weight_temperature`	0.5
`loss_weight_refresh_interval`	50
`stop_on_divergence`	True
`divergence_gap`	0.002
`divergence_ratio`	1.5
`divergence_patience`	50
`divergence_min_updates`	10
`val_spike_threshold`	2.0
`val_spike_window`	15
`val_spike_frequency`	0.75
`val_plateau_patience`	250
`val_plateau_min_delta`	0.0001
`custom_lr`	0.0001
`disable_lr_scaling`	True
`custom_warmup`	-1
`lr_min_ratio`	0.001
`resume_warmup_ratio`	0.05
`plateau_factor`	0.8
`plateau_patience`	15
`preserve_optimizer`	False
`preserve_scheduler`	True
`samples_mode`	Train additional samples
`num_random_obs_to_visualize`	2
`selected_frame_offset`	3
`runs_per_stage`	5
`serial_runs`	True
`clean_old_checkpoints`	True
`enable_baseline`	False
`baseline_runs_per_stage`	1
`run_id`	shoulder_session_multiheight_vgg_loss_2
`enable_wandb`	True
`wandb_project`	developmental-robot-movement
`lr_sweep.lr_min`	1e-07
`lr_sweep.lr_max`	0.01
`lr_sweep.phase_a_num_candidates`	5
`lr_sweep.phase_a_seeds`	1
`lr_sweep.phase_a_time_budget_min`	3.0
`lr_sweep.phase_a_survivor_count`	2
`lr_sweep.phase_b_seeds`	3
`lr_sweep.phase_b_time_budget_min`	5.0
`lr_sweep.ranking_metric`	median_best_val
`lr_sweep.min_samples_before_timeout`	1000
`lr_sweep.min_evals_before_stop`	5
`lr_sweep.save_sweep_state`	True
`plateau_sweep.enabled`	True
`plateau_sweep.plateau_ema_alpha`	0.95
`plateau_sweep.plateau_improvement_threshold`	0.005
`plateau_sweep.plateau_patience`	25
`plateau_sweep.cooldown_updates`	5
`plateau_sweep.max_sweeps_per_stage`	2
`plateau_sweep.min_sweep_improvement`	0.0
`initial_sweep_enabled`	True
`stage_time_budget_min`	180

Timing Summary

Stage	Plateau Sweeps	Sweep Time	Training Time	Stage Total
Stage 1	2	00:49:50	00:07:10	00:57:01
TOTAL	2	00:49:50	00:07:10	00:57:01

Initial LR Sweep: Stage 1: selected LR 3.16e-05 in 00:22:35

Plateau Sweep Details

Total Sweeps: 2
Stages with Sweeps: 1 of 1
Total Sweep Time: 00:49:50
Average Sweep Duration: 00:24:55

Stage 1: 2 sweeps

LR Progression: 3.2e-05 → 1.0e-07 → 1.0e-07

Sweep #	Triggered At (samples)	Wall Time	Selected LR	Duration
1	16,128	00:03:24	1.00e-07	00:24:04
2	25,856	00:29:32	1.00e-07	00:25:46

Stage Results

Stage	Best Loss	Stop Reason	Samples Trained	Time	Sweeps	LR (Initial→Final)
Stage 1	0.119720	max_sweeps (2)	7,936	00:57:01	2	3.2e-05→1.0e-07

Total Plateau Sweeps: 2

Stop Reason Breakdown

max_sweeps: 1 stages

Loss Across Full Training Run

Multi-Run Statistics

Total Runs: 5
Average Best Loss: 0.120076 ± 0.000208
Best Overall: 0.119720
Worst Overall: 0.120235

Stage 1 (5 runs)

Run	Best Loss	Stop Reason	Samples	Time	Selected
1	N/A	sweep_no_improvement (val: 0.122552 -> 0.123502, delta: -0.000950)	7,680	00:52:22
2	0.120204	max_sweeps (2)	9,728	01:28:15
3	0.120143	max_sweeps (2)	12,288	00:34:58
4	0.120235	max_sweeps (2)	8,960	01:14:03
5	0.119720	max_sweeps (2)	7,936	00:57:01	✓
Mean: 0.120076 ± 0.000208		Min: 0.119720 / Max: 0.120235		Range: 0.000515

Best Checkpoint

Name: best_model_auto_session_so101_multiheight_part1_1345_shoulder_session_multiheight_vgg_loss_2_00027392_cont_val_0.119720.pth
Stage: 1
Hybrid Loss (full session): 0.130517

Learning Rate Timeline with Plateau Sweeps

Stage Progression

Stage	Orig Loss	Train Loss	Time	Samples	Stop Reason
1 ⭐	0.130517	0.119720	00:57:01	7936	max_sweeps (2)

Hybrid Loss Over Original Session (per Stage)

Stage 1 (Best) - Hybrid Loss: 0.130517

Sample Counts

Cumulative Across All Stages

Per Stage

Stage 1 (Best) - Total Samples: 7,936

Best Checkpoint Inference

Selected Frame 3

Action 0

Action 1

Action 2

Random Observations

Observation 248

Action 0

Action 1

Action 2

Observation 1167

Action 0

Action 1

Action 2

Individual Stage Reports

Stage 1 Run 5 (Staged) (orig loss: 0.130517) ⭐